x86: dsputil_mmx: Remove leftover inline assembly fragments
[ffmpeg.git] / libavcodec / x86 / dsputil_mmx.c
1 /*
2  * MMX optimized DSP utils
3  * Copyright (c) 2000, 2001 Fabrice Bellard
4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * MMX optimization by Nick Kurshev <nickols_k@mail.ru>
7  *
8  * This file is part of Libav.
9  *
10  * Libav is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * Libav is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with Libav; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23  */
24
25 #include "libavutil/attributes.h"
26 #include "libavutil/cpu.h"
27 #include "libavutil/x86/asm.h"
28 #include "libavcodec/dsputil.h"
29 #include "libavcodec/h264dsp.h"
30 #include "libavcodec/mpegvideo.h"
31 #include "libavcodec/simple_idct.h"
32 #include "dsputil_mmx.h"
33 #include "idct_xvid.h"
34
35 //#undef NDEBUG
36 //#include <assert.h>
37
38 /* pixel operations */
39 DECLARE_ALIGNED(8,  const uint64_t, ff_bone) = 0x0101010101010101ULL;
40 DECLARE_ALIGNED(8,  const uint64_t, ff_wtwo) = 0x0002000200020002ULL;
41
42 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1)    = { 0x0001000100010001ULL, 0x0001000100010001ULL };
43 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_2)    = { 0x0002000200020002ULL, 0x0002000200020002ULL };
44 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_3)    = { 0x0003000300030003ULL, 0x0003000300030003ULL };
45 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_4)    = { 0x0004000400040004ULL, 0x0004000400040004ULL };
46 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_5)    = { 0x0005000500050005ULL, 0x0005000500050005ULL };
47 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_8)    = { 0x0008000800080008ULL, 0x0008000800080008ULL };
48 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_9)    = { 0x0009000900090009ULL, 0x0009000900090009ULL };
49 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_15)   =   0x000F000F000F000FULL;
50 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_16)   = { 0x0010001000100010ULL, 0x0010001000100010ULL };
51 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_17)   = { 0x0011001100110011ULL, 0x0011001100110011ULL };
52 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_18)   = { 0x0012001200120012ULL, 0x0012001200120012ULL };
53 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_20)   =   0x0014001400140014ULL;
54 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_27)   = { 0x001B001B001B001BULL, 0x001B001B001B001BULL };
55 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_28)   = { 0x001C001C001C001CULL, 0x001C001C001C001CULL };
56 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_32)   = { 0x0020002000200020ULL, 0x0020002000200020ULL };
57 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_42)   =   0x002A002A002A002AULL;
58 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_53)   =   0x0035003500350035ULL;
59 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_63)   = { 0x003F003F003F003FULL, 0x003F003F003F003FULL };
60 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x0040004000400040ULL };
61 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
62 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
63 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
64 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
65 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
66
67 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_0)    = { 0x0000000000000000ULL, 0x0000000000000000ULL };
68 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_1)    = { 0x0101010101010101ULL, 0x0101010101010101ULL };
69 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_3)    = { 0x0303030303030303ULL, 0x0303030303030303ULL };
70 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_4)    = { 0x0404040404040404ULL, 0x0404040404040404ULL };
71 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_7)    =   0x0707070707070707ULL;
72 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_1F)   =   0x1F1F1F1F1F1F1F1FULL;
73 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_3F)   =   0x3F3F3F3F3F3F3F3FULL;
74 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_80)   = { 0x8080808080808080ULL, 0x8080808080808080ULL };
75 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_81)   =   0x8181818181818181ULL;
76 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_A1)   = { 0xA1A1A1A1A1A1A1A1ULL, 0xA1A1A1A1A1A1A1A1ULL };
77 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_F8)   = { 0xF8F8F8F8F8F8F8F8ULL, 0xF8F8F8F8F8F8F8F8ULL };
78 DECLARE_ALIGNED(8,  const uint64_t, ff_pb_FC)   =   0xFCFCFCFCFCFCFCFCULL;
79 DECLARE_ALIGNED(16, const xmm_reg,  ff_pb_FE)   = { 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL };
80
81 DECLARE_ALIGNED(16, const double, ff_pd_1)[2] = { 1.0, 1.0 };
82 DECLARE_ALIGNED(16, const double, ff_pd_2)[2] = { 2.0, 2.0 };
83
84
85 #if HAVE_YASM
86 void ff_put_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
87                               ptrdiff_t line_size, int h);
88 void ff_put_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
89                              ptrdiff_t line_size, int h);
90 void ff_put_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
91                               int dstStride, int src1Stride, int h);
92 void ff_put_no_rnd_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1,
93                                      uint8_t *src2, int dstStride,
94                                      int src1Stride, int h);
95 void ff_avg_pixels8_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
96                               int dstStride, int src1Stride, int h);
97 void ff_put_pixels16_x2_mmxext(uint8_t *block, const uint8_t *pixels,
98                                ptrdiff_t line_size, int h);
99 void ff_put_pixels16_x2_3dnow(uint8_t *block, const uint8_t *pixels,
100                               ptrdiff_t line_size, int h);
101 void ff_put_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
102                                int dstStride, int src1Stride, int h);
103 void ff_avg_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
104                                int dstStride, int src1Stride, int h);
105 void ff_put_no_rnd_pixels16_l2_mmxext(uint8_t *dst, uint8_t *src1, uint8_t *src2,
106                                       int dstStride, int src1Stride, int h);
107 void ff_put_no_rnd_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
108                                      ptrdiff_t line_size, int h);
109 void ff_put_no_rnd_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
110                                     ptrdiff_t line_size, int h);
111 void ff_put_no_rnd_pixels8_x2_exact_mmxext(uint8_t *block,
112                                            const uint8_t *pixels,
113                                            ptrdiff_t line_size, int h);
114 void ff_put_no_rnd_pixels8_x2_exact_3dnow(uint8_t *block,
115                                           const uint8_t *pixels,
116                                           ptrdiff_t line_size, int h);
117 void ff_put_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
118                               ptrdiff_t line_size, int h);
119 void ff_put_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
120                              ptrdiff_t line_size, int h);
121 void ff_put_no_rnd_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
122                                      ptrdiff_t line_size, int h);
123 void ff_put_no_rnd_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
124                                     ptrdiff_t line_size, int h);
125 void ff_put_no_rnd_pixels8_y2_exact_mmxext(uint8_t *block,
126                                            const uint8_t *pixels,
127                                            ptrdiff_t line_size, int h);
128 void ff_put_no_rnd_pixels8_y2_exact_3dnow(uint8_t *block,
129                                           const uint8_t *pixels,
130                                           ptrdiff_t line_size, int h);
131 void ff_avg_pixels8_mmxext(uint8_t *block, const uint8_t *pixels,
132                            ptrdiff_t line_size, int h);
133 void ff_avg_pixels8_3dnow(uint8_t *block, const uint8_t *pixels,
134                           ptrdiff_t line_size, int h);
135 void ff_avg_pixels8_x2_mmxext(uint8_t *block, const uint8_t *pixels,
136                               ptrdiff_t line_size, int h);
137 void ff_avg_pixels8_x2_3dnow(uint8_t *block, const uint8_t *pixels,
138                              ptrdiff_t line_size, int h);
139 void ff_avg_pixels8_y2_mmxext(uint8_t *block, const uint8_t *pixels,
140                               ptrdiff_t line_size, int h);
141 void ff_avg_pixels8_y2_3dnow(uint8_t *block, const uint8_t *pixels,
142                              ptrdiff_t line_size, int h);
143 void ff_avg_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
144                                ptrdiff_t line_size, int h);
145 void ff_avg_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
146                               ptrdiff_t line_size, int h);
147
148 void ff_put_pixels8_mmxext(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h);
149 static void ff_put_pixels16_mmxext(uint8_t *block, const uint8_t *pixels,
150                                    ptrdiff_t line_size, int h)
151 {
152     ff_put_pixels8_mmxext(block,     pixels,     line_size, h);
153     ff_put_pixels8_mmxext(block + 8, pixels + 8, line_size, h);
154 }
155
156 void ff_put_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
157                                          int dstStride, int srcStride, int h);
158 void ff_avg_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
159                                          int dstStride, int srcStride, int h);
160 void ff_put_no_rnd_mpeg4_qpel16_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
161                                                  int dstStride, int srcStride,
162                                                  int h);
163 void ff_put_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
164                                         int dstStride, int srcStride, int h);
165 void ff_avg_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
166                                         int dstStride, int srcStride, int h);
167 void ff_put_no_rnd_mpeg4_qpel8_h_lowpass_mmxext(uint8_t *dst, uint8_t *src,
168                                                 int dstStride, int srcStride,
169                                                 int h);
170 void ff_put_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
171                                          int dstStride, int srcStride);
172 void ff_avg_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
173                                          int dstStride, int srcStride);
174 void ff_put_no_rnd_mpeg4_qpel16_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
175                                                  int dstStride, int srcStride);
176 void ff_put_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
177                                         int dstStride, int srcStride);
178 void ff_avg_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
179                                         int dstStride, int srcStride);
180 void ff_put_no_rnd_mpeg4_qpel8_v_lowpass_mmxext(uint8_t *dst, uint8_t *src,
181                                                 int dstStride, int srcStride);
182 #define ff_put_no_rnd_pixels16_mmxext ff_put_pixels16_mmxext
183 #define ff_put_no_rnd_pixels8_mmxext ff_put_pixels8_mmxext
184 #endif /* HAVE_YASM */
185
186
187 #if HAVE_INLINE_ASM
188
189 #define JUMPALIGN()     __asm__ volatile (".p2align 3"::)
190 #define MOVQ_ZERO(regd) __asm__ volatile ("pxor %%"#regd", %%"#regd ::)
191
192 #define MOVQ_BFE(regd)                                  \
193     __asm__ volatile (                                  \
194         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
195         "paddb   %%"#regd", %%"#regd"   \n\t" ::)
196
197 #ifndef PIC
198 #define MOVQ_BONE(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_bone))
199 #define MOVQ_WTWO(regd) __asm__ volatile ("movq %0, %%"#regd" \n\t" :: "m"(ff_wtwo))
200 #else
201 // for shared library it's better to use this way for accessing constants
202 // pcmpeqd -> -1
203 #define MOVQ_BONE(regd)                                 \
204     __asm__ volatile (                                  \
205         "pcmpeqd  %%"#regd", %%"#regd"  \n\t"           \
206         "psrlw          $15, %%"#regd"  \n\t"           \
207         "packuswb %%"#regd", %%"#regd"  \n\t" ::)
208
209 #define MOVQ_WTWO(regd)                                 \
210     __asm__ volatile (                                  \
211         "pcmpeqd %%"#regd", %%"#regd"   \n\t"           \
212         "psrlw         $15, %%"#regd"   \n\t"           \
213         "psllw          $1, %%"#regd"   \n\t"::)
214
215 #endif
216
217 // using regr as temporary and for the output result
218 // first argument is unmodifed and second is trashed
219 // regfe is supposed to contain 0xfefefefefefefefe
220 #define PAVGB_MMX_NO_RND(rega, regb, regr, regfe)                \
221     "movq   "#rega", "#regr"            \n\t"                    \
222     "pand   "#regb", "#regr"            \n\t"                    \
223     "pxor   "#rega", "#regb"            \n\t"                    \
224     "pand  "#regfe", "#regb"            \n\t"                    \
225     "psrlq       $1, "#regb"            \n\t"                    \
226     "paddb  "#regb", "#regr"            \n\t"
227
228 #define PAVGB_MMX(rega, regb, regr, regfe)                       \
229     "movq   "#rega", "#regr"            \n\t"                    \
230     "por    "#regb", "#regr"            \n\t"                    \
231     "pxor   "#rega", "#regb"            \n\t"                    \
232     "pand  "#regfe", "#regb"            \n\t"                    \
233     "psrlq       $1, "#regb"            \n\t"                    \
234     "psubb  "#regb", "#regr"            \n\t"
235
236 // mm6 is supposed to contain 0xfefefefefefefefe
237 #define PAVGBP_MMX_NO_RND(rega, regb, regr,  regc, regd, regp)   \
238     "movq  "#rega", "#regr"             \n\t"                    \
239     "movq  "#regc", "#regp"             \n\t"                    \
240     "pand  "#regb", "#regr"             \n\t"                    \
241     "pand  "#regd", "#regp"             \n\t"                    \
242     "pxor  "#rega", "#regb"             \n\t"                    \
243     "pxor  "#regc", "#regd"             \n\t"                    \
244     "pand    %%mm6, "#regb"             \n\t"                    \
245     "pand    %%mm6, "#regd"             \n\t"                    \
246     "psrlq      $1, "#regb"             \n\t"                    \
247     "psrlq      $1, "#regd"             \n\t"                    \
248     "paddb "#regb", "#regr"             \n\t"                    \
249     "paddb "#regd", "#regp"             \n\t"
250
251 #define PAVGBP_MMX(rega, regb, regr, regc, regd, regp)           \
252     "movq  "#rega", "#regr"             \n\t"                    \
253     "movq  "#regc", "#regp"             \n\t"                    \
254     "por   "#regb", "#regr"             \n\t"                    \
255     "por   "#regd", "#regp"             \n\t"                    \
256     "pxor  "#rega", "#regb"             \n\t"                    \
257     "pxor  "#regc", "#regd"             \n\t"                    \
258     "pand    %%mm6, "#regb"             \n\t"                    \
259     "pand    %%mm6, "#regd"             \n\t"                    \
260     "psrlq      $1, "#regd"             \n\t"                    \
261     "psrlq      $1, "#regb"             \n\t"                    \
262     "psubb "#regb", "#regr"             \n\t"                    \
263     "psubb "#regd", "#regp"             \n\t"
264
265 /***********************************/
266 /* MMX no rounding */
267 #define NO_RND 1
268 #define DEF(x, y) x ## _no_rnd_ ## y ## _mmx
269 #define SET_RND  MOVQ_WONE
270 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX_NO_RND(a, b, c, d, e, f)
271 #define PAVGB(a, b, c, e)               PAVGB_MMX_NO_RND(a, b, c, e)
272 #define OP_AVG(a, b, c, e)              PAVGB_MMX(a, b, c, e)
273
274 #include "dsputil_rnd_template.c"
275
276 #undef DEF
277 #undef SET_RND
278 #undef PAVGBP
279 #undef PAVGB
280 #undef NO_RND
281 /***********************************/
282 /* MMX rounding */
283
284 #define DEF(x, y) x ## _ ## y ## _mmx
285 #define SET_RND  MOVQ_WTWO
286 #define PAVGBP(a, b, c, d, e, f)        PAVGBP_MMX(a, b, c, d, e, f)
287 #define PAVGB(a, b, c, e)               PAVGB_MMX(a, b, c, e)
288
289 #include "dsputil_rnd_template.c"
290
291 #undef DEF
292 #undef SET_RND
293 #undef PAVGBP
294 #undef PAVGB
295 #undef OP_AVG
296
297 #endif /* HAVE_INLINE_ASM */
298
299
300 #if HAVE_YASM
301 #define ff_put_pixels8_mmx ff_put_pixels8_mmxext
302
303 /***********************************/
304 /* 3Dnow specific */
305
306 #define DEF(x) x ## _3dnow
307
308 #include "dsputil_avg_template.c"
309
310 #undef DEF
311
312 /***********************************/
313 /* MMXEXT specific */
314
315 #define DEF(x) x ## _mmxext
316
317 #include "dsputil_avg_template.c"
318
319 #undef DEF
320
321 #endif /* HAVE_YASM */
322
323
324 #if HAVE_INLINE_ASM
325 #define put_no_rnd_pixels16_mmx put_pixels16_mmx
326 #define put_no_rnd_pixels8_mmx put_pixels8_mmx
327 #define put_pixels16_mmxext put_pixels16_mmx
328 #define put_pixels8_mmxext put_pixels8_mmx
329 #define put_pixels4_mmxext put_pixels4_mmx
330 #define put_no_rnd_pixels16_mmxext put_no_rnd_pixels16_mmx
331 #define put_no_rnd_pixels8_mmxext put_no_rnd_pixels8_mmx
332
333 /***********************************/
334 /* standard MMX */
335
336 void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
337                                int line_size)
338 {
339     const int16_t *p;
340     uint8_t *pix;
341
342     /* read the pixels */
343     p   = block;
344     pix = pixels;
345     /* unrolled loop */
346     __asm__ volatile (
347         "movq      (%3), %%mm0          \n\t"
348         "movq     8(%3), %%mm1          \n\t"
349         "movq    16(%3), %%mm2          \n\t"
350         "movq    24(%3), %%mm3          \n\t"
351         "movq    32(%3), %%mm4          \n\t"
352         "movq    40(%3), %%mm5          \n\t"
353         "movq    48(%3), %%mm6          \n\t"
354         "movq    56(%3), %%mm7          \n\t"
355         "packuswb %%mm1, %%mm0          \n\t"
356         "packuswb %%mm3, %%mm2          \n\t"
357         "packuswb %%mm5, %%mm4          \n\t"
358         "packuswb %%mm7, %%mm6          \n\t"
359         "movq     %%mm0, (%0)           \n\t"
360         "movq     %%mm2, (%0, %1)       \n\t"
361         "movq     %%mm4, (%0, %1, 2)    \n\t"
362         "movq     %%mm6, (%0, %2)       \n\t"
363         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3),
364            "r"(p)
365         : "memory");
366     pix += line_size * 4;
367     p   += 32;
368
369     // if here would be an exact copy of the code above
370     // compiler would generate some very strange code
371     // thus using "r"
372     __asm__ volatile (
373         "movq       (%3), %%mm0         \n\t"
374         "movq      8(%3), %%mm1         \n\t"
375         "movq     16(%3), %%mm2         \n\t"
376         "movq     24(%3), %%mm3         \n\t"
377         "movq     32(%3), %%mm4         \n\t"
378         "movq     40(%3), %%mm5         \n\t"
379         "movq     48(%3), %%mm6         \n\t"
380         "movq     56(%3), %%mm7         \n\t"
381         "packuswb  %%mm1, %%mm0         \n\t"
382         "packuswb  %%mm3, %%mm2         \n\t"
383         "packuswb  %%mm5, %%mm4         \n\t"
384         "packuswb  %%mm7, %%mm6         \n\t"
385         "movq      %%mm0, (%0)          \n\t"
386         "movq      %%mm2, (%0, %1)      \n\t"
387         "movq      %%mm4, (%0, %1, 2)   \n\t"
388         "movq      %%mm6, (%0, %2)      \n\t"
389         :: "r"(pix), "r"((x86_reg)line_size), "r"((x86_reg)line_size * 3), "r"(p)
390         : "memory");
391 }
392
393 #define put_signed_pixels_clamped_mmx_half(off)             \
394     "movq          "#off"(%2), %%mm1        \n\t"           \
395     "movq     16 + "#off"(%2), %%mm2        \n\t"           \
396     "movq     32 + "#off"(%2), %%mm3        \n\t"           \
397     "movq     48 + "#off"(%2), %%mm4        \n\t"           \
398     "packsswb  8 + "#off"(%2), %%mm1        \n\t"           \
399     "packsswb 24 + "#off"(%2), %%mm2        \n\t"           \
400     "packsswb 40 + "#off"(%2), %%mm3        \n\t"           \
401     "packsswb 56 + "#off"(%2), %%mm4        \n\t"           \
402     "paddb              %%mm0, %%mm1        \n\t"           \
403     "paddb              %%mm0, %%mm2        \n\t"           \
404     "paddb              %%mm0, %%mm3        \n\t"           \
405     "paddb              %%mm0, %%mm4        \n\t"           \
406     "movq               %%mm1, (%0)         \n\t"           \
407     "movq               %%mm2, (%0, %3)     \n\t"           \
408     "movq               %%mm3, (%0, %3, 2)  \n\t"           \
409     "movq               %%mm4, (%0, %1)     \n\t"
410
411 void ff_put_signed_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
412                                       int line_size)
413 {
414     x86_reg line_skip = line_size;
415     x86_reg line_skip3;
416
417     __asm__ volatile (
418         "movq "MANGLE(ff_pb_80)", %%mm0     \n\t"
419         "lea         (%3, %3, 2), %1        \n\t"
420         put_signed_pixels_clamped_mmx_half(0)
421         "lea         (%0, %3, 4), %0        \n\t"
422         put_signed_pixels_clamped_mmx_half(64)
423         : "+&r"(pixels), "=&r"(line_skip3)
424         : "r"(block), "r"(line_skip)
425         : "memory");
426 }
427
428 void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels,
429                                int line_size)
430 {
431     const int16_t *p;
432     uint8_t *pix;
433     int i;
434
435     /* read the pixels */
436     p   = block;
437     pix = pixels;
438     MOVQ_ZERO(mm7);
439     i = 4;
440     do {
441         __asm__ volatile (
442             "movq        (%2), %%mm0    \n\t"
443             "movq       8(%2), %%mm1    \n\t"
444             "movq      16(%2), %%mm2    \n\t"
445             "movq      24(%2), %%mm3    \n\t"
446             "movq          %0, %%mm4    \n\t"
447             "movq          %1, %%mm6    \n\t"
448             "movq       %%mm4, %%mm5    \n\t"
449             "punpcklbw  %%mm7, %%mm4    \n\t"
450             "punpckhbw  %%mm7, %%mm5    \n\t"
451             "paddsw     %%mm4, %%mm0    \n\t"
452             "paddsw     %%mm5, %%mm1    \n\t"
453             "movq       %%mm6, %%mm5    \n\t"
454             "punpcklbw  %%mm7, %%mm6    \n\t"
455             "punpckhbw  %%mm7, %%mm5    \n\t"
456             "paddsw     %%mm6, %%mm2    \n\t"
457             "paddsw     %%mm5, %%mm3    \n\t"
458             "packuswb   %%mm1, %%mm0    \n\t"
459             "packuswb   %%mm3, %%mm2    \n\t"
460             "movq       %%mm0, %0       \n\t"
461             "movq       %%mm2, %1       \n\t"
462             : "+m"(*pix), "+m"(*(pix + line_size))
463             : "r"(p)
464             : "memory");
465         pix += line_size * 2;
466         p   += 16;
467     } while (--i);
468 }
469
470 static void put_pixels8_mmx(uint8_t *block, const uint8_t *pixels,
471                             int line_size, int h)
472 {
473     __asm__ volatile (
474         "lea   (%3, %3), %%"REG_a"      \n\t"
475         ".p2align     3                 \n\t"
476         "1:                             \n\t"
477         "movq  (%1    ), %%mm0          \n\t"
478         "movq  (%1, %3), %%mm1          \n\t"
479         "movq     %%mm0, (%2)           \n\t"
480         "movq     %%mm1, (%2, %3)       \n\t"
481         "add  %%"REG_a", %1             \n\t"
482         "add  %%"REG_a", %2             \n\t"
483         "movq  (%1    ), %%mm0          \n\t"
484         "movq  (%1, %3), %%mm1          \n\t"
485         "movq     %%mm0, (%2)           \n\t"
486         "movq     %%mm1, (%2, %3)       \n\t"
487         "add  %%"REG_a", %1             \n\t"
488         "add  %%"REG_a", %2             \n\t"
489         "subl        $4, %0             \n\t"
490         "jnz         1b                 \n\t"
491         : "+g"(h), "+r"(pixels),  "+r"(block)
492         : "r"((x86_reg)line_size)
493         : "%"REG_a, "memory"
494         );
495 }
496
497 static void put_pixels16_mmx(uint8_t *block, const uint8_t *pixels,
498                              int line_size, int h)
499 {
500     __asm__ volatile (
501         "lea   (%3, %3), %%"REG_a"      \n\t"
502         ".p2align     3                 \n\t"
503         "1:                             \n\t"
504         "movq  (%1    ), %%mm0          \n\t"
505         "movq 8(%1    ), %%mm4          \n\t"
506         "movq  (%1, %3), %%mm1          \n\t"
507         "movq 8(%1, %3), %%mm5          \n\t"
508         "movq     %%mm0,  (%2)          \n\t"
509         "movq     %%mm4, 8(%2)          \n\t"
510         "movq     %%mm1,  (%2, %3)      \n\t"
511         "movq     %%mm5, 8(%2, %3)      \n\t"
512         "add  %%"REG_a", %1             \n\t"
513         "add  %%"REG_a", %2             \n\t"
514         "movq  (%1    ), %%mm0          \n\t"
515         "movq 8(%1    ), %%mm4          \n\t"
516         "movq  (%1, %3), %%mm1          \n\t"
517         "movq 8(%1, %3), %%mm5          \n\t"
518         "movq     %%mm0,  (%2)          \n\t"
519         "movq     %%mm4, 8(%2)          \n\t"
520         "movq     %%mm1,  (%2, %3)      \n\t"
521         "movq     %%mm5, 8(%2, %3)      \n\t"
522         "add  %%"REG_a", %1             \n\t"
523         "add  %%"REG_a", %2             \n\t"
524         "subl        $4, %0             \n\t"
525         "jnz         1b                 \n\t"
526         : "+g"(h), "+r"(pixels),  "+r"(block)
527         : "r"((x86_reg)line_size)
528         : "%"REG_a, "memory"
529         );
530 }
531
532 #define CLEAR_BLOCKS(name, n)                           \
533 static void name(int16_t *blocks)                       \
534 {                                                       \
535     __asm__ volatile (                                  \
536         "pxor %%mm7, %%mm7              \n\t"           \
537         "mov     %1,        %%"REG_a"   \n\t"           \
538         "1:                             \n\t"           \
539         "movq %%mm7,   (%0, %%"REG_a")  \n\t"           \
540         "movq %%mm7,  8(%0, %%"REG_a")  \n\t"           \
541         "movq %%mm7, 16(%0, %%"REG_a")  \n\t"           \
542         "movq %%mm7, 24(%0, %%"REG_a")  \n\t"           \
543         "add    $32, %%"REG_a"          \n\t"           \
544         "js      1b                     \n\t"           \
545         :: "r"(((uint8_t *)blocks) + 128 * n),          \
546            "i"(-128 * n)                                \
547         : "%"REG_a                                      \
548         );                                              \
549 }
550 CLEAR_BLOCKS(clear_blocks_mmx, 6)
551 CLEAR_BLOCKS(clear_block_mmx, 1)
552
553 static void clear_block_sse(int16_t *block)
554 {
555     __asm__ volatile (
556         "xorps  %%xmm0, %%xmm0          \n"
557         "movaps %%xmm0,    (%0)         \n"
558         "movaps %%xmm0,  16(%0)         \n"
559         "movaps %%xmm0,  32(%0)         \n"
560         "movaps %%xmm0,  48(%0)         \n"
561         "movaps %%xmm0,  64(%0)         \n"
562         "movaps %%xmm0,  80(%0)         \n"
563         "movaps %%xmm0,  96(%0)         \n"
564         "movaps %%xmm0, 112(%0)         \n"
565         :: "r"(block)
566         : "memory"
567     );
568 }
569
570 static void clear_blocks_sse(int16_t *blocks)
571 {
572     __asm__ volatile (
573         "xorps  %%xmm0, %%xmm0              \n"
574         "mov        %1,         %%"REG_a"   \n"
575         "1:                                 \n"
576         "movaps %%xmm0,    (%0, %%"REG_a")  \n"
577         "movaps %%xmm0,  16(%0, %%"REG_a")  \n"
578         "movaps %%xmm0,  32(%0, %%"REG_a")  \n"
579         "movaps %%xmm0,  48(%0, %%"REG_a")  \n"
580         "movaps %%xmm0,  64(%0, %%"REG_a")  \n"
581         "movaps %%xmm0,  80(%0, %%"REG_a")  \n"
582         "movaps %%xmm0,  96(%0, %%"REG_a")  \n"
583         "movaps %%xmm0, 112(%0, %%"REG_a")  \n"
584         "add      $128,         %%"REG_a"   \n"
585         "js         1b                      \n"
586         :: "r"(((uint8_t *)blocks) + 128 * 6),
587            "i"(-128 * 6)
588         : "%"REG_a
589     );
590 }
591
592 static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w)
593 {
594     x86_reg i = 0;
595     __asm__ volatile (
596         "jmp          2f                \n\t"
597         "1:                             \n\t"
598         "movq   (%1, %0), %%mm0         \n\t"
599         "movq   (%2, %0), %%mm1         \n\t"
600         "paddb     %%mm0, %%mm1         \n\t"
601         "movq      %%mm1, (%2, %0)      \n\t"
602         "movq  8(%1, %0), %%mm0         \n\t"
603         "movq  8(%2, %0), %%mm1         \n\t"
604         "paddb     %%mm0, %%mm1         \n\t"
605         "movq      %%mm1, 8(%2, %0)     \n\t"
606         "add         $16, %0            \n\t"
607         "2:                             \n\t"
608         "cmp          %3, %0            \n\t"
609         "js           1b                \n\t"
610         : "+r"(i)
611         : "r"(src), "r"(dst), "r"((x86_reg)w - 15)
612     );
613     for ( ; i < w; i++)
614         dst[i + 0] += src[i + 0];
615 }
616
617 #if HAVE_7REGS
618 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top,
619                                             const uint8_t *diff, int w,
620                                             int *left, int *left_top)
621 {
622     x86_reg w2 = -w;
623     x86_reg x;
624     int l  = *left     & 0xff;
625     int tl = *left_top & 0xff;
626     int t;
627     __asm__ volatile (
628         "mov          %7, %3            \n"
629         "1:                             \n"
630         "movzbl (%3, %4), %2            \n"
631         "mov          %2, %k3           \n"
632         "sub         %b1, %b3           \n"
633         "add         %b0, %b3           \n"
634         "mov          %2, %1            \n"
635         "cmp          %0, %2            \n"
636         "cmovg        %0, %2            \n"
637         "cmovg        %1, %0            \n"
638         "cmp         %k3, %0            \n"
639         "cmovg       %k3, %0            \n"
640         "mov          %7, %3            \n"
641         "cmp          %2, %0            \n"
642         "cmovl        %2, %0            \n"
643         "add    (%6, %4), %b0           \n"
644         "mov         %b0, (%5, %4)      \n"
645         "inc          %4                \n"
646         "jl           1b                \n"
647         : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
648         : "r"(dst + w), "r"(diff + w), "rm"(top + w)
649     );
650     *left     = l;
651     *left_top = tl;
652 }
653 #endif
654 #endif /* HAVE_INLINE_ASM */
655
656 void ff_h263_v_loop_filter_mmx(uint8_t *src, int stride, int qscale);
657 void ff_h263_h_loop_filter_mmx(uint8_t *src, int stride, int qscale);
658
659 #if HAVE_INLINE_ASM
660 /* Draw the edges of width 'w' of an image of size width, height
661  * this MMX version can only handle w == 8 || w == 16. */
662 static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height,
663                            int w, int h, int sides)
664 {
665     uint8_t *ptr, *last_line;
666     int i;
667
668     last_line = buf + (height - 1) * wrap;
669     /* left and right */
670     ptr = buf;
671     if (w == 8) {
672         __asm__ volatile (
673             "1:                             \n\t"
674             "movd            (%0), %%mm0    \n\t"
675             "punpcklbw      %%mm0, %%mm0    \n\t"
676             "punpcklwd      %%mm0, %%mm0    \n\t"
677             "punpckldq      %%mm0, %%mm0    \n\t"
678             "movq           %%mm0, -8(%0)   \n\t"
679             "movq      -8(%0, %2), %%mm1    \n\t"
680             "punpckhbw      %%mm1, %%mm1    \n\t"
681             "punpckhwd      %%mm1, %%mm1    \n\t"
682             "punpckhdq      %%mm1, %%mm1    \n\t"
683             "movq           %%mm1, (%0, %2) \n\t"
684             "add               %1, %0       \n\t"
685             "cmp               %3, %0       \n\t"
686             "jb                1b           \n\t"
687             : "+r"(ptr)
688             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
689             );
690     } else {
691         __asm__ volatile (
692             "1:                                 \n\t"
693             "movd            (%0), %%mm0        \n\t"
694             "punpcklbw      %%mm0, %%mm0        \n\t"
695             "punpcklwd      %%mm0, %%mm0        \n\t"
696             "punpckldq      %%mm0, %%mm0        \n\t"
697             "movq           %%mm0, -8(%0)       \n\t"
698             "movq           %%mm0, -16(%0)      \n\t"
699             "movq      -8(%0, %2), %%mm1        \n\t"
700             "punpckhbw      %%mm1, %%mm1        \n\t"
701             "punpckhwd      %%mm1, %%mm1        \n\t"
702             "punpckhdq      %%mm1, %%mm1        \n\t"
703             "movq           %%mm1,  (%0, %2)    \n\t"
704             "movq           %%mm1, 8(%0, %2)    \n\t"
705             "add               %1, %0           \n\t"
706             "cmp               %3, %0           \n\t"
707             "jb                1b               \n\t"
708             : "+r"(ptr)
709             : "r"((x86_reg)wrap), "r"((x86_reg)width), "r"(ptr + wrap * height)
710             );
711     }
712
713     /* top and bottom (and hopefully also the corners) */
714     if (sides & EDGE_TOP) {
715         for (i = 0; i < h; i += 4) {
716             ptr = buf - (i + 1) * wrap - w;
717             __asm__ volatile (
718                 "1:                             \n\t"
719                 "movq (%1, %0), %%mm0           \n\t"
720                 "movq    %%mm0, (%0)            \n\t"
721                 "movq    %%mm0, (%0, %2)        \n\t"
722                 "movq    %%mm0, (%0, %2, 2)     \n\t"
723                 "movq    %%mm0, (%0, %3)        \n\t"
724                 "add        $8, %0              \n\t"
725                 "cmp        %4, %0              \n\t"
726                 "jb         1b                  \n\t"
727                 : "+r"(ptr)
728                 : "r"((x86_reg)buf - (x86_reg)ptr - w), "r"((x86_reg) -wrap),
729                   "r"((x86_reg) -wrap * 3), "r"(ptr + width + 2 * w)
730                 );
731         }
732     }
733
734     if (sides & EDGE_BOTTOM) {
735         for (i = 0; i < h; i += 4) {
736             ptr = last_line + (i + 1) * wrap - w;
737             __asm__ volatile (
738                 "1:                             \n\t"
739                 "movq (%1, %0), %%mm0           \n\t"
740                 "movq    %%mm0, (%0)            \n\t"
741                 "movq    %%mm0, (%0, %2)        \n\t"
742                 "movq    %%mm0, (%0, %2, 2)     \n\t"
743                 "movq    %%mm0, (%0, %3)        \n\t"
744                 "add        $8, %0              \n\t"
745                 "cmp        %4, %0              \n\t"
746                 "jb         1b                  \n\t"
747                 : "+r"(ptr)
748                 : "r"((x86_reg)last_line - (x86_reg)ptr - w),
749                   "r"((x86_reg)wrap), "r"((x86_reg)wrap * 3),
750                   "r"(ptr + width + 2 * w)
751                 );
752         }
753     }
754 }
755 #endif /* HAVE_INLINE_ASM */
756
757
758 #if HAVE_YASM
759 #define QPEL_OP(OPNAME, ROUNDER, RND, MMX)                              \
760 static void OPNAME ## qpel8_mc00_ ## MMX (uint8_t *dst, uint8_t *src,   \
761                                           int stride)                   \
762 {                                                                       \
763     ff_ ## OPNAME ## pixels8_ ## MMX(dst, src, stride, 8);              \
764 }                                                                       \
765                                                                         \
766 static void OPNAME ## qpel8_mc10_ ## MMX(uint8_t *dst, uint8_t *src,    \
767                                          int stride)                    \
768 {                                                                       \
769     uint64_t temp[8];                                                   \
770     uint8_t * const half = (uint8_t*)temp;                              \
771     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
772                                                    stride, 8);          \
773     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
774                                         stride, stride, 8);             \
775 }                                                                       \
776                                                                         \
777 static void OPNAME ## qpel8_mc20_ ## MMX(uint8_t *dst, uint8_t *src,    \
778                                          int stride)                    \
779 {                                                                       \
780     ff_ ## OPNAME ## mpeg4_qpel8_h_lowpass_ ## MMX(dst, src, stride,    \
781                                                    stride, 8);          \
782 }                                                                       \
783                                                                         \
784 static void OPNAME ## qpel8_mc30_ ## MMX(uint8_t *dst, uint8_t *src,    \
785                                          int stride)                    \
786 {                                                                       \
787     uint64_t temp[8];                                                   \
788     uint8_t * const half = (uint8_t*)temp;                              \
789     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(half, src, 8,        \
790                                                    stride, 8);          \
791     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + 1, half, stride,     \
792                                         stride, 8);                     \
793 }                                                                       \
794                                                                         \
795 static void OPNAME ## qpel8_mc01_ ## MMX(uint8_t *dst, uint8_t *src,    \
796                                          int stride)                    \
797 {                                                                       \
798     uint64_t temp[8];                                                   \
799     uint8_t * const half = (uint8_t*)temp;                              \
800     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
801                                                    8, stride);          \
802     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src, half,                 \
803                                         stride, stride, 8);             \
804 }                                                                       \
805                                                                         \
806 static void OPNAME ## qpel8_mc02_ ## MMX(uint8_t *dst, uint8_t *src,    \
807                                          int stride)                    \
808 {                                                                       \
809     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, src,            \
810                                                    stride, stride);     \
811 }                                                                       \
812                                                                         \
813 static void OPNAME ## qpel8_mc03_ ## MMX(uint8_t *dst, uint8_t *src,    \
814                                          int stride)                    \
815 {                                                                       \
816     uint64_t temp[8];                                                   \
817     uint8_t * const half = (uint8_t*)temp;                              \
818     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(half, src,           \
819                                                    8, stride);          \
820     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, src + stride, half, stride,\
821                                         stride, 8);                     \
822 }                                                                       \
823                                                                         \
824 static void OPNAME ## qpel8_mc11_ ## MMX(uint8_t *dst, uint8_t *src,    \
825                                          int stride)                    \
826 {                                                                       \
827     uint64_t half[8 + 9];                                               \
828     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
829     uint8_t * const halfHV = ((uint8_t*)half);                          \
830     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
831                                                    stride, 9);          \
832     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
833                                         stride, 9);                     \
834     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
835     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
836                                         stride, 8, 8);                  \
837 }                                                                       \
838                                                                         \
839 static void OPNAME ## qpel8_mc31_ ## MMX(uint8_t *dst, uint8_t *src,    \
840                                          int stride)                    \
841 {                                                                       \
842     uint64_t half[8 + 9];                                               \
843     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
844     uint8_t * const halfHV = ((uint8_t*)half);                          \
845     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
846                                                    stride, 9);          \
847     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
848                                         stride, 9);                     \
849     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
850     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
851                                         stride, 8, 8);                  \
852 }                                                                       \
853                                                                         \
854 static void OPNAME ## qpel8_mc13_ ## MMX(uint8_t *dst, uint8_t *src,    \
855                                          int stride)                    \
856 {                                                                       \
857     uint64_t half[8 + 9];                                               \
858     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
859     uint8_t * const halfHV = ((uint8_t*)half);                          \
860     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
861                                                    stride, 9);          \
862     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH, 8,           \
863                                         stride, 9);                     \
864     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
865     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
866                                         stride, 8, 8);                  \
867 }                                                                       \
868                                                                         \
869 static void OPNAME ## qpel8_mc33_ ## MMX(uint8_t *dst, uint8_t *src,    \
870                                          int stride)                    \
871 {                                                                       \
872     uint64_t half[8 + 9];                                               \
873     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
874     uint8_t * const halfHV = ((uint8_t*)half);                          \
875     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
876                                                    stride, 9);          \
877     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
878                                         stride, 9);                     \
879     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
880     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
881                                         stride, 8, 8);                  \
882 }                                                                       \
883                                                                         \
884 static void OPNAME ## qpel8_mc21_ ## MMX(uint8_t *dst, uint8_t *src,    \
885                                          int stride)                    \
886 {                                                                       \
887     uint64_t half[8 + 9];                                               \
888     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
889     uint8_t * const halfHV = ((uint8_t*)half);                          \
890     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
891                                                    stride, 9);          \
892     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
893     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH, halfHV,             \
894                                         stride, 8, 8);                  \
895 }                                                                       \
896                                                                         \
897 static void OPNAME ## qpel8_mc23_ ## MMX(uint8_t *dst, uint8_t *src,    \
898                                          int stride)                    \
899 {                                                                       \
900     uint64_t half[8 + 9];                                               \
901     uint8_t * const halfH  = ((uint8_t*)half) + 64;                     \
902     uint8_t * const halfHV = ((uint8_t*)half);                          \
903     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
904                                                    stride, 9);          \
905     ff_put ## RND ## mpeg4_qpel8_v_lowpass_ ## MMX(halfHV, halfH, 8, 8);\
906     ff_ ## OPNAME ## pixels8_l2_ ## MMX(dst, halfH + 8, halfHV,         \
907                                         stride, 8, 8);                  \
908 }                                                                       \
909                                                                         \
910 static void OPNAME ## qpel8_mc12_ ## MMX(uint8_t *dst, uint8_t *src,    \
911                                          int stride)                    \
912 {                                                                       \
913     uint64_t half[8 + 9];                                               \
914     uint8_t * const halfH = ((uint8_t*)half);                           \
915     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
916                                                    stride, 9);          \
917     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src, halfH,              \
918                                         8, stride, 9);                  \
919     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
920                                                    stride, 8);          \
921 }                                                                       \
922                                                                         \
923 static void OPNAME ## qpel8_mc32_ ## MMX(uint8_t *dst, uint8_t *src,    \
924                                          int stride)                    \
925 {                                                                       \
926     uint64_t half[8 + 9];                                               \
927     uint8_t * const halfH = ((uint8_t*)half);                           \
928     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
929                                                    stride, 9);          \
930     ff_put ## RND ## pixels8_l2_ ## MMX(halfH, src + 1, halfH, 8,       \
931                                         stride, 9);                     \
932     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
933                                                    stride, 8);          \
934 }                                                                       \
935                                                                         \
936 static void OPNAME ## qpel8_mc22_ ## MMX(uint8_t *dst, uint8_t *src,    \
937                                          int stride)                    \
938 {                                                                       \
939     uint64_t half[9];                                                   \
940     uint8_t * const halfH = ((uint8_t*)half);                           \
941     ff_put ## RND ## mpeg4_qpel8_h_lowpass_ ## MMX(halfH, src, 8,       \
942                                                    stride, 9);          \
943     ff_ ## OPNAME ## mpeg4_qpel8_v_lowpass_ ## MMX(dst, halfH,          \
944                                                    stride, 8);          \
945 }                                                                       \
946                                                                         \
947 static void OPNAME ## qpel16_mc00_ ## MMX (uint8_t *dst, uint8_t *src,  \
948                                            int stride)                  \
949 {                                                                       \
950     ff_ ## OPNAME ## pixels16_ ## MMX(dst, src, stride, 16);            \
951 }                                                                       \
952                                                                         \
953 static void OPNAME ## qpel16_mc10_ ## MMX(uint8_t *dst, uint8_t *src,   \
954                                           int stride)                   \
955 {                                                                       \
956     uint64_t temp[32];                                                  \
957     uint8_t * const half = (uint8_t*)temp;                              \
958     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
959                                                     stride, 16);        \
960     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
961                                          stride, 16);                   \
962 }                                                                       \
963                                                                         \
964 static void OPNAME ## qpel16_mc20_ ## MMX(uint8_t *dst, uint8_t *src,   \
965                                           int stride)                   \
966 {                                                                       \
967     ff_ ## OPNAME ## mpeg4_qpel16_h_lowpass_ ## MMX(dst, src,           \
968                                                     stride, stride, 16);\
969 }                                                                       \
970                                                                         \
971 static void OPNAME ## qpel16_mc30_ ## MMX(uint8_t *dst, uint8_t *src,   \
972                                           int stride)                   \
973 {                                                                       \
974     uint64_t temp[32];                                                  \
975     uint8_t * const half = (uint8_t*)temp;                              \
976     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(half, src, 16,      \
977                                                     stride, 16);        \
978     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src + 1, half,            \
979                                          stride, stride, 16);           \
980 }                                                                       \
981                                                                         \
982 static void OPNAME ## qpel16_mc01_ ## MMX(uint8_t *dst, uint8_t *src,   \
983                                           int stride)                   \
984 {                                                                       \
985     uint64_t temp[32];                                                  \
986     uint8_t * const half = (uint8_t*)temp;                              \
987     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
988                                                     stride);            \
989     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src, half, stride,        \
990                                          stride, 16);                   \
991 }                                                                       \
992                                                                         \
993 static void OPNAME ## qpel16_mc02_ ## MMX(uint8_t *dst, uint8_t *src,   \
994                                           int stride)                   \
995 {                                                                       \
996     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, src,           \
997                                                     stride, stride);    \
998 }                                                                       \
999                                                                         \
1000 static void OPNAME ## qpel16_mc03_ ## MMX(uint8_t *dst, uint8_t *src,   \
1001                                           int stride)                   \
1002 {                                                                       \
1003     uint64_t temp[32];                                                  \
1004     uint8_t * const half = (uint8_t*)temp;                              \
1005     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(half, src, 16,      \
1006                                                     stride);            \
1007     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, src+stride, half,         \
1008                                          stride, stride, 16);           \
1009 }                                                                       \
1010                                                                         \
1011 static void OPNAME ## qpel16_mc11_ ## MMX(uint8_t *dst, uint8_t *src,   \
1012                                           int stride)                   \
1013 {                                                                       \
1014     uint64_t half[16 * 2 + 17 * 2];                                     \
1015     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1016     uint8_t * const halfHV = ((uint8_t*)half);                          \
1017     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1018                                                     stride, 17);        \
1019     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1020                                          stride, 17);                   \
1021     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1022                                                     16, 16);            \
1023     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1024                                          stride, 16, 16);               \
1025 }                                                                       \
1026                                                                         \
1027 static void OPNAME ## qpel16_mc31_ ## MMX(uint8_t *dst, uint8_t *src,   \
1028                                           int stride)                   \
1029 {                                                                       \
1030     uint64_t half[16 * 2 + 17 * 2];                                     \
1031     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1032     uint8_t * const halfHV = ((uint8_t*)half);                          \
1033     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1034                                                     stride, 17);        \
1035     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1036                                          stride, 17);                   \
1037     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1038                                                     16, 16);            \
1039     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1040                                          stride, 16, 16);               \
1041 }                                                                       \
1042                                                                         \
1043 static void OPNAME ## qpel16_mc13_ ## MMX(uint8_t *dst, uint8_t *src,   \
1044                                           int stride)                   \
1045 {                                                                       \
1046     uint64_t half[16 * 2 + 17 * 2];                                     \
1047     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1048     uint8_t * const halfHV = ((uint8_t*)half);                          \
1049     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1050                                                     stride, 17);        \
1051     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1052                                          stride, 17);                   \
1053     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1054                                                     16, 16);            \
1055     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1056                                          stride, 16, 16);               \
1057 }                                                                       \
1058                                                                         \
1059 static void OPNAME ## qpel16_mc33_ ## MMX(uint8_t *dst, uint8_t *src,   \
1060                                           int stride)                   \
1061 {                                                                       \
1062     uint64_t half[16 * 2 + 17 * 2];                                     \
1063     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1064     uint8_t * const halfHV = ((uint8_t*)half);                          \
1065     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1066                                                     stride, 17);        \
1067     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1068                                          stride, 17);                   \
1069     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1070                                                     16, 16);            \
1071     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1072                                          stride, 16, 16);               \
1073 }                                                                       \
1074                                                                         \
1075 static void OPNAME ## qpel16_mc21_ ## MMX(uint8_t *dst, uint8_t *src,   \
1076                                           int stride)                   \
1077 {                                                                       \
1078     uint64_t half[16 * 2 + 17 * 2];                                     \
1079     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1080     uint8_t * const halfHV = ((uint8_t*)half);                          \
1081     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1082                                                     stride, 17);        \
1083     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1084                                                     16, 16);            \
1085     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH, halfHV,            \
1086                                          stride, 16, 16);               \
1087 }                                                                       \
1088                                                                         \
1089 static void OPNAME ## qpel16_mc23_ ## MMX(uint8_t *dst, uint8_t *src,   \
1090                                           int stride)                   \
1091 {                                                                       \
1092     uint64_t half[16 * 2 + 17 * 2];                                     \
1093     uint8_t * const halfH  = ((uint8_t*)half) + 256;                    \
1094     uint8_t * const halfHV = ((uint8_t*)half);                          \
1095     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1096                                                     stride, 17);        \
1097     ff_put ## RND ## mpeg4_qpel16_v_lowpass_ ## MMX(halfHV, halfH,      \
1098                                                     16, 16);            \
1099     ff_ ## OPNAME ## pixels16_l2_ ## MMX(dst, halfH + 16, halfHV,       \
1100                                          stride, 16, 16);               \
1101 }                                                                       \
1102                                                                         \
1103 static void OPNAME ## qpel16_mc12_ ## MMX(uint8_t *dst, uint8_t *src,   \
1104                                           int stride)                   \
1105 {                                                                       \
1106     uint64_t half[17 * 2];                                              \
1107     uint8_t * const halfH = ((uint8_t*)half);                           \
1108     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1109                                                     stride, 17);        \
1110     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src, halfH, 16,         \
1111                                          stride, 17);                   \
1112     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1113                                                     stride, 16);        \
1114 }                                                                       \
1115                                                                         \
1116 static void OPNAME ## qpel16_mc32_ ## MMX(uint8_t *dst, uint8_t *src,   \
1117                                           int stride)                   \
1118 {                                                                       \
1119     uint64_t half[17 * 2];                                              \
1120     uint8_t * const halfH = ((uint8_t*)half);                           \
1121     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1122                                                     stride, 17);        \
1123     ff_put ## RND ## pixels16_l2_ ## MMX(halfH, src + 1, halfH, 16,     \
1124                                          stride, 17);                   \
1125     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1126                                                     stride, 16);        \
1127 }                                                                       \
1128                                                                         \
1129 static void OPNAME ## qpel16_mc22_ ## MMX(uint8_t *dst, uint8_t *src,   \
1130                                           int stride)                   \
1131 {                                                                       \
1132     uint64_t half[17 * 2];                                              \
1133     uint8_t * const halfH = ((uint8_t*)half);                           \
1134     ff_put ## RND ## mpeg4_qpel16_h_lowpass_ ## MMX(halfH, src, 16,     \
1135                                                     stride, 17);        \
1136     ff_ ## OPNAME ## mpeg4_qpel16_v_lowpass_ ## MMX(dst, halfH,         \
1137                                                     stride, 16);        \
1138 }
1139
1140 QPEL_OP(put_,          ff_pw_16, _,        mmxext)
1141 QPEL_OP(avg_,          ff_pw_16, _,        mmxext)
1142 QPEL_OP(put_no_rnd_,   ff_pw_15, _no_rnd_, mmxext)
1143 #endif /* HAVE_YASM */
1144
1145
1146 #if HAVE_INLINE_ASM
1147 void ff_put_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1148 {
1149   put_pixels8_xy2_mmx(dst, src, stride, 8);
1150 }
1151 void ff_put_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1152 {
1153   put_pixels16_xy2_mmx(dst, src, stride, 16);
1154 }
1155 void ff_avg_rv40_qpel8_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1156 {
1157   avg_pixels8_xy2_mmx(dst, src, stride, 8);
1158 }
1159 void ff_avg_rv40_qpel16_mc33_mmx(uint8_t *dst, uint8_t *src, int stride)
1160 {
1161   avg_pixels16_xy2_mmx(dst, src, stride, 16);
1162 }
1163
1164 static void gmc_mmx(uint8_t *dst, uint8_t *src,
1165                     int stride, int h, int ox, int oy,
1166                     int dxx, int dxy, int dyx, int dyy,
1167                     int shift, int r, int width, int height)
1168 {
1169     const int w    = 8;
1170     const int ix   = ox  >> (16 + shift);
1171     const int iy   = oy  >> (16 + shift);
1172     const int oxs  = ox  >> 4;
1173     const int oys  = oy  >> 4;
1174     const int dxxs = dxx >> 4;
1175     const int dxys = dxy >> 4;
1176     const int dyxs = dyx >> 4;
1177     const int dyys = dyy >> 4;
1178     const uint16_t r4[4]   = { r, r, r, r };
1179     const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
1180     const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
1181     const uint64_t shift2 = 2 * shift;
1182     int x, y;
1183
1184     const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
1185     const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
1186     const int dxh = dxy * (h - 1);
1187     const int dyw = dyx * (w - 1);
1188     if ( // non-constant fullpel offset (3% of blocks)
1189         ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
1190          (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift)
1191         // uses more than 16 bits of subpel mv (only at huge resolution)
1192         || (dxx | dxy | dyx | dyy) & 15 ||
1193         (unsigned)ix >= width  - w ||
1194         (unsigned)iy >= height - h) {
1195         // FIXME could still use mmx for some of the rows
1196         ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
1197                  shift, r, width, height);
1198         return;
1199     }
1200
1201     src += ix + iy * stride;
1202
1203     __asm__ volatile (
1204         "movd         %0, %%mm6         \n\t"
1205         "pxor      %%mm7, %%mm7         \n\t"
1206         "punpcklwd %%mm6, %%mm6         \n\t"
1207         "punpcklwd %%mm6, %%mm6         \n\t"
1208         :: "r"(1<<shift)
1209     );
1210
1211     for (x = 0; x < w; x += 4) {
1212         uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
1213                             oxs - dxys + dxxs * (x + 1),
1214                             oxs - dxys + dxxs * (x + 2),
1215                             oxs - dxys + dxxs * (x + 3) };
1216         uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
1217                             oys - dyys + dyxs * (x + 1),
1218                             oys - dyys + dyxs * (x + 2),
1219                             oys - dyys + dyxs * (x + 3) };
1220
1221         for (y = 0; y < h; y++) {
1222             __asm__ volatile (
1223                 "movq      %0, %%mm4    \n\t"
1224                 "movq      %1, %%mm5    \n\t"
1225                 "paddw     %2, %%mm4    \n\t"
1226                 "paddw     %3, %%mm5    \n\t"
1227                 "movq   %%mm4, %0       \n\t"
1228                 "movq   %%mm5, %1       \n\t"
1229                 "psrlw    $12, %%mm4    \n\t"
1230                 "psrlw    $12, %%mm5    \n\t"
1231                 : "+m"(*dx4), "+m"(*dy4)
1232                 : "m"(*dxy4), "m"(*dyy4)
1233             );
1234
1235             __asm__ volatile (
1236                 "movq      %%mm6, %%mm2 \n\t"
1237                 "movq      %%mm6, %%mm1 \n\t"
1238                 "psubw     %%mm4, %%mm2 \n\t"
1239                 "psubw     %%mm5, %%mm1 \n\t"
1240                 "movq      %%mm2, %%mm0 \n\t"
1241                 "movq      %%mm4, %%mm3 \n\t"
1242                 "pmullw    %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
1243                 "pmullw    %%mm5, %%mm3 \n\t" // dx * dy
1244                 "pmullw    %%mm5, %%mm2 \n\t" // (s - dx) * dy
1245                 "pmullw    %%mm4, %%mm1 \n\t" // dx * (s - dy)
1246
1247                 "movd         %4, %%mm5 \n\t"
1248                 "movd         %3, %%mm4 \n\t"
1249                 "punpcklbw %%mm7, %%mm5 \n\t"
1250                 "punpcklbw %%mm7, %%mm4 \n\t"
1251                 "pmullw    %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
1252                 "pmullw    %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
1253
1254                 "movd         %2, %%mm5 \n\t"
1255                 "movd         %1, %%mm4 \n\t"
1256                 "punpcklbw %%mm7, %%mm5 \n\t"
1257                 "punpcklbw %%mm7, %%mm4 \n\t"
1258                 "pmullw    %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
1259                 "pmullw    %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
1260                 "paddw        %5, %%mm1 \n\t"
1261                 "paddw     %%mm3, %%mm2 \n\t"
1262                 "paddw     %%mm1, %%mm0 \n\t"
1263                 "paddw     %%mm2, %%mm0 \n\t"
1264
1265                 "psrlw        %6, %%mm0 \n\t"
1266                 "packuswb  %%mm0, %%mm0 \n\t"
1267                 "movd      %%mm0, %0    \n\t"
1268
1269                 : "=m"(dst[x + y * stride])
1270                 : "m"(src[0]), "m"(src[1]),
1271                   "m"(src[stride]), "m"(src[stride + 1]),
1272                   "m"(*r4), "m"(shift2)
1273             );
1274             src += stride;
1275         }
1276         src += 4 - h * stride;
1277     }
1278 }
1279 #endif /* HAVE_INLINE_ASM */
1280
1281 void ff_put_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1282                           ptrdiff_t line_size, int h);
1283 void ff_avg_pixels16_sse2(uint8_t *block, const uint8_t *pixels,
1284                           ptrdiff_t line_size, int h);
1285
1286 #if HAVE_INLINE_ASM
1287
1288 /* CAVS-specific */
1289 void ff_put_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1290 {
1291     put_pixels8_mmx(dst, src, stride, 8);
1292 }
1293
1294 void ff_avg_cavs_qpel8_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1295 {
1296     avg_pixels8_mmx(dst, src, stride, 8);
1297 }
1298
1299 void ff_put_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1300 {
1301     put_pixels16_mmx(dst, src, stride, 16);
1302 }
1303
1304 void ff_avg_cavs_qpel16_mc00_mmxext(uint8_t *dst, uint8_t *src, int stride)
1305 {
1306     avg_pixels16_mmx(dst, src, stride, 16);
1307 }
1308 #endif /* HAVE_INLINE_ASM */
1309
1310 #if HAVE_YASM
1311 /* VC-1-specific */
1312 void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src,
1313                                int stride, int rnd)
1314 {
1315     ff_put_pixels8_mmx(dst, src, stride, 8);
1316 }
1317
1318 void ff_avg_vc1_mspel_mc00_mmxext(uint8_t *dst, const uint8_t *src,
1319                                   int stride, int rnd)
1320 {
1321     ff_avg_pixels8_mmxext(dst, src, stride, 8);
1322 }
1323 #endif /* HAVE_YASM */
1324
1325 #if HAVE_INLINE_ASM
1326 static void vector_clipf_sse(float *dst, const float *src,
1327                              float min, float max, int len)
1328 {
1329     x86_reg i = (len - 16) * 4;
1330     __asm__ volatile (
1331         "movss          %3, %%xmm4      \n\t"
1332         "movss          %4, %%xmm5      \n\t"
1333         "shufps $0, %%xmm4, %%xmm4      \n\t"
1334         "shufps $0, %%xmm5, %%xmm5      \n\t"
1335         "1:                             \n\t"
1336         "movaps   (%2, %0), %%xmm0      \n\t" // 3/1 on intel
1337         "movaps 16(%2, %0), %%xmm1      \n\t"
1338         "movaps 32(%2, %0), %%xmm2      \n\t"
1339         "movaps 48(%2, %0), %%xmm3      \n\t"
1340         "maxps      %%xmm4, %%xmm0      \n\t"
1341         "maxps      %%xmm4, %%xmm1      \n\t"
1342         "maxps      %%xmm4, %%xmm2      \n\t"
1343         "maxps      %%xmm4, %%xmm3      \n\t"
1344         "minps      %%xmm5, %%xmm0      \n\t"
1345         "minps      %%xmm5, %%xmm1      \n\t"
1346         "minps      %%xmm5, %%xmm2      \n\t"
1347         "minps      %%xmm5, %%xmm3      \n\t"
1348         "movaps     %%xmm0,   (%1, %0)  \n\t"
1349         "movaps     %%xmm1, 16(%1, %0)  \n\t"
1350         "movaps     %%xmm2, 32(%1, %0)  \n\t"
1351         "movaps     %%xmm3, 48(%1, %0)  \n\t"
1352         "sub           $64, %0          \n\t"
1353         "jge            1b              \n\t"
1354         : "+&r"(i)
1355         : "r"(dst), "r"(src), "m"(min), "m"(max)
1356         : "memory"
1357     );
1358 }
1359
1360 #endif /* HAVE_INLINE_ASM */
1361
1362 int32_t ff_scalarproduct_int16_mmxext(const int16_t *v1, const int16_t *v2,
1363                                       int order);
1364 int32_t ff_scalarproduct_int16_sse2(const int16_t *v1, const int16_t *v2,
1365                                     int order);
1366 int32_t ff_scalarproduct_and_madd_int16_mmxext(int16_t *v1, const int16_t *v2,
1367                                                const int16_t *v3,
1368                                                int order, int mul);
1369 int32_t ff_scalarproduct_and_madd_int16_sse2(int16_t *v1, const int16_t *v2,
1370                                              const int16_t *v3,
1371                                              int order, int mul);
1372 int32_t ff_scalarproduct_and_madd_int16_ssse3(int16_t *v1, const int16_t *v2,
1373                                               const int16_t *v3,
1374                                               int order, int mul);
1375
1376 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
1377                                         const int16_t *window, unsigned int len);
1378 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
1379                                       const int16_t *window, unsigned int len);
1380 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
1381                                   const int16_t *window, unsigned int len);
1382 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
1383                                 const int16_t *window, unsigned int len);
1384 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
1385                                  const int16_t *window, unsigned int len);
1386 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
1387                                       const int16_t *window, unsigned int len);
1388
1389 void ff_bswap32_buf_ssse3(uint32_t *dst, const uint32_t *src, int w);
1390 void ff_bswap32_buf_sse2(uint32_t *dst, const uint32_t *src, int w);
1391
1392 void ff_add_hfyu_median_prediction_mmxext(uint8_t *dst, const uint8_t *top,
1393                                           const uint8_t *diff, int w,
1394                                           int *left, int *left_top);
1395 int  ff_add_hfyu_left_prediction_ssse3(uint8_t *dst, const uint8_t *src,
1396                                        int w, int left);
1397 int  ff_add_hfyu_left_prediction_sse4(uint8_t *dst, const uint8_t *src,
1398                                       int w, int left);
1399
1400 void ff_vector_clip_int32_mmx     (int32_t *dst, const int32_t *src,
1401                                    int32_t min, int32_t max, unsigned int len);
1402 void ff_vector_clip_int32_sse2    (int32_t *dst, const int32_t *src,
1403                                    int32_t min, int32_t max, unsigned int len);
1404 void ff_vector_clip_int32_int_sse2(int32_t *dst, const int32_t *src,
1405                                    int32_t min, int32_t max, unsigned int len);
1406 void ff_vector_clip_int32_sse4    (int32_t *dst, const int32_t *src,
1407                                    int32_t min, int32_t max, unsigned int len);
1408
1409 #define SET_QPEL_FUNCS(PFX, IDX, SIZE, CPU, PREFIX)                          \
1410     do {                                                                     \
1411     c->PFX ## _pixels_tab[IDX][ 0] = PREFIX ## PFX ## SIZE ## _mc00_ ## CPU; \
1412     c->PFX ## _pixels_tab[IDX][ 1] = PREFIX ## PFX ## SIZE ## _mc10_ ## CPU; \
1413     c->PFX ## _pixels_tab[IDX][ 2] = PREFIX ## PFX ## SIZE ## _mc20_ ## CPU; \
1414     c->PFX ## _pixels_tab[IDX][ 3] = PREFIX ## PFX ## SIZE ## _mc30_ ## CPU; \
1415     c->PFX ## _pixels_tab[IDX][ 4] = PREFIX ## PFX ## SIZE ## _mc01_ ## CPU; \
1416     c->PFX ## _pixels_tab[IDX][ 5] = PREFIX ## PFX ## SIZE ## _mc11_ ## CPU; \
1417     c->PFX ## _pixels_tab[IDX][ 6] = PREFIX ## PFX ## SIZE ## _mc21_ ## CPU; \
1418     c->PFX ## _pixels_tab[IDX][ 7] = PREFIX ## PFX ## SIZE ## _mc31_ ## CPU; \
1419     c->PFX ## _pixels_tab[IDX][ 8] = PREFIX ## PFX ## SIZE ## _mc02_ ## CPU; \
1420     c->PFX ## _pixels_tab[IDX][ 9] = PREFIX ## PFX ## SIZE ## _mc12_ ## CPU; \
1421     c->PFX ## _pixels_tab[IDX][10] = PREFIX ## PFX ## SIZE ## _mc22_ ## CPU; \
1422     c->PFX ## _pixels_tab[IDX][11] = PREFIX ## PFX ## SIZE ## _mc32_ ## CPU; \
1423     c->PFX ## _pixels_tab[IDX][12] = PREFIX ## PFX ## SIZE ## _mc03_ ## CPU; \
1424     c->PFX ## _pixels_tab[IDX][13] = PREFIX ## PFX ## SIZE ## _mc13_ ## CPU; \
1425     c->PFX ## _pixels_tab[IDX][14] = PREFIX ## PFX ## SIZE ## _mc23_ ## CPU; \
1426     c->PFX ## _pixels_tab[IDX][15] = PREFIX ## PFX ## SIZE ## _mc33_ ## CPU; \
1427     } while (0)
1428
1429 #define SET_HPEL_FUNCS(PFX, IDX, SIZE, CPU)                                     \
1430     do {                                                                        \
1431         c->PFX ## _pixels_tab IDX [0] = PFX ## _pixels ## SIZE ## _     ## CPU; \
1432         c->PFX ## _pixels_tab IDX [1] = PFX ## _pixels ## SIZE ## _x2_  ## CPU; \
1433         c->PFX ## _pixels_tab IDX [2] = PFX ## _pixels ## SIZE ## _y2_  ## CPU; \
1434         c->PFX ## _pixels_tab IDX [3] = PFX ## _pixels ## SIZE ## _xy2_ ## CPU; \
1435     } while (0)
1436
1437 static av_cold void dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx,
1438                                      int mm_flags)
1439 {
1440     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1441
1442 #if HAVE_INLINE_ASM
1443     c->put_pixels_clamped        = ff_put_pixels_clamped_mmx;
1444     c->put_signed_pixels_clamped = ff_put_signed_pixels_clamped_mmx;
1445     c->add_pixels_clamped        = ff_add_pixels_clamped_mmx;
1446
1447     if (!high_bit_depth) {
1448         c->clear_block  = clear_block_mmx;
1449         c->clear_blocks = clear_blocks_mmx;
1450         c->draw_edges   = draw_edges_mmx;
1451
1452         SET_HPEL_FUNCS(put,        [0], 16, mmx);
1453         SET_HPEL_FUNCS(put_no_rnd, [0], 16, mmx);
1454         SET_HPEL_FUNCS(avg,        [0], 16, mmx);
1455         SET_HPEL_FUNCS(avg_no_rnd,    , 16, mmx);
1456         SET_HPEL_FUNCS(put,        [1],  8, mmx);
1457         SET_HPEL_FUNCS(put_no_rnd, [1],  8, mmx);
1458         SET_HPEL_FUNCS(avg,        [1],  8, mmx);
1459
1460         switch (avctx->idct_algo) {
1461         case FF_IDCT_AUTO:
1462         case FF_IDCT_SIMPLEMMX:
1463             c->idct_put              = ff_simple_idct_put_mmx;
1464             c->idct_add              = ff_simple_idct_add_mmx;
1465             c->idct                  = ff_simple_idct_mmx;
1466             c->idct_permutation_type = FF_SIMPLE_IDCT_PERM;
1467             break;
1468         case FF_IDCT_XVIDMMX:
1469             c->idct_put              = ff_idct_xvid_mmx_put;
1470             c->idct_add              = ff_idct_xvid_mmx_add;
1471             c->idct                  = ff_idct_xvid_mmx;
1472             break;
1473         }
1474     }
1475
1476     c->gmc = gmc_mmx;
1477
1478     c->add_bytes = add_bytes_mmx;
1479 #endif /* HAVE_INLINE_ASM */
1480
1481 #if HAVE_YASM
1482     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
1483         c->h263_v_loop_filter = ff_h263_v_loop_filter_mmx;
1484         c->h263_h_loop_filter = ff_h263_h_loop_filter_mmx;
1485     }
1486
1487     c->vector_clip_int32 = ff_vector_clip_int32_mmx;
1488 #endif
1489
1490 }
1491
1492 static av_cold void dsputil_init_mmxext(DSPContext *c, AVCodecContext *avctx,
1493                                         int mm_flags)
1494 {
1495     const int bit_depth      = avctx->bits_per_raw_sample;
1496     const int high_bit_depth = bit_depth > 8;
1497
1498 #if HAVE_YASM
1499     SET_QPEL_FUNCS(avg_qpel,        0, 16, mmxext, );
1500     SET_QPEL_FUNCS(avg_qpel,        1,  8, mmxext, );
1501
1502     SET_QPEL_FUNCS(put_qpel,        0, 16, mmxext, );
1503     SET_QPEL_FUNCS(put_qpel,        1,  8, mmxext, );
1504     SET_QPEL_FUNCS(put_no_rnd_qpel, 0, 16, mmxext, );
1505     SET_QPEL_FUNCS(put_no_rnd_qpel, 1,  8, mmxext, );
1506
1507     if (!high_bit_depth) {
1508         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_mmxext;
1509         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_mmxext;
1510
1511         c->avg_pixels_tab[0][0] = ff_avg_pixels16_mmxext;
1512         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_mmxext;
1513         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_mmxext;
1514
1515         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_mmxext;
1516         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_mmxext;
1517
1518         c->avg_pixels_tab[1][0] = ff_avg_pixels8_mmxext;
1519         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_mmxext;
1520         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_mmxext;
1521     }
1522
1523     if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
1524         if (!high_bit_depth) {
1525             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_mmxext;
1526             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_mmxext;
1527             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_mmxext;
1528             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_mmxext;
1529
1530             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_mmxext;
1531             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_mmxext;
1532         }
1533     }
1534 #endif /* HAVE_YASM */
1535
1536 #if HAVE_INLINE_ASM
1537     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1538         c->idct_put = ff_idct_xvid_mmxext_put;
1539         c->idct_add = ff_idct_xvid_mmxext_add;
1540         c->idct     = ff_idct_xvid_mmxext;
1541     }
1542 #endif /* HAVE_INLINE_ASM */
1543
1544 #if HAVE_MMXEXT_EXTERNAL
1545     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1546                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1547         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_mmxext;
1548         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_mmxext;
1549     }
1550
1551     /* slower than cmov version on AMD */
1552     if (!(mm_flags & AV_CPU_FLAG_3DNOW))
1553         c->add_hfyu_median_prediction = ff_add_hfyu_median_prediction_mmxext;
1554
1555     c->scalarproduct_int16          = ff_scalarproduct_int16_mmxext;
1556     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_mmxext;
1557
1558     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1559         c->apply_window_int16 = ff_apply_window_int16_mmxext;
1560     } else {
1561         c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
1562     }
1563 #endif /* HAVE_MMXEXT_EXTERNAL */
1564 }
1565
1566 static av_cold void dsputil_init_3dnow(DSPContext *c, AVCodecContext *avctx,
1567                                        int mm_flags)
1568 {
1569     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1570
1571 #if HAVE_YASM
1572     if (!high_bit_depth) {
1573         c->put_pixels_tab[0][1] = ff_put_pixels16_x2_3dnow;
1574         c->put_pixels_tab[0][2] = ff_put_pixels16_y2_3dnow;
1575
1576         c->avg_pixels_tab[0][0] = ff_avg_pixels16_3dnow;
1577         c->avg_pixels_tab[0][1] = ff_avg_pixels16_x2_3dnow;
1578         c->avg_pixels_tab[0][2] = ff_avg_pixels16_y2_3dnow;
1579
1580         c->put_pixels_tab[1][1] = ff_put_pixels8_x2_3dnow;
1581         c->put_pixels_tab[1][2] = ff_put_pixels8_y2_3dnow;
1582
1583         c->avg_pixels_tab[1][0] = ff_avg_pixels8_3dnow;
1584         c->avg_pixels_tab[1][1] = ff_avg_pixels8_x2_3dnow;
1585         c->avg_pixels_tab[1][2] = ff_avg_pixels8_y2_3dnow;
1586
1587         if (!(avctx->flags & CODEC_FLAG_BITEXACT)){
1588             c->put_no_rnd_pixels_tab[0][1] = ff_put_no_rnd_pixels16_x2_3dnow;
1589             c->put_no_rnd_pixels_tab[0][2] = ff_put_no_rnd_pixels16_y2_3dnow;
1590             c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_3dnow;
1591             c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_3dnow;
1592
1593             c->avg_pixels_tab[0][3] = ff_avg_pixels16_xy2_3dnow;
1594             c->avg_pixels_tab[1][3] = ff_avg_pixels8_xy2_3dnow;
1595         }
1596     }
1597
1598     if (CONFIG_VP3_DECODER && (avctx->codec_id == AV_CODEC_ID_VP3 ||
1599                                avctx->codec_id == AV_CODEC_ID_THEORA)) {
1600         c->put_no_rnd_pixels_tab[1][1] = ff_put_no_rnd_pixels8_x2_exact_3dnow;
1601         c->put_no_rnd_pixels_tab[1][2] = ff_put_no_rnd_pixels8_y2_exact_3dnow;
1602     }
1603 #endif /* HAVE_YASM */
1604 }
1605
1606 static av_cold void dsputil_init_sse(DSPContext *c, AVCodecContext *avctx,
1607                                      int mm_flags)
1608 {
1609     const int high_bit_depth = avctx->bits_per_raw_sample > 8;
1610
1611 #if HAVE_INLINE_ASM
1612     if (!high_bit_depth) {
1613         if (!(CONFIG_MPEG_XVMC_DECODER && avctx->xvmc_acceleration > 1)) {
1614             /* XvMCCreateBlocks() may not allocate 16-byte aligned blocks */
1615             c->clear_block  = clear_block_sse;
1616             c->clear_blocks = clear_blocks_sse;
1617         }
1618     }
1619
1620     c->vector_clipf = vector_clipf_sse;
1621 #endif /* HAVE_INLINE_ASM */
1622 }
1623
1624 static av_cold void dsputil_init_sse2(DSPContext *c, AVCodecContext *avctx,
1625                                       int mm_flags)
1626 {
1627     const int bit_depth      = avctx->bits_per_raw_sample;
1628     const int high_bit_depth = bit_depth > 8;
1629
1630 #if HAVE_SSE2_INLINE
1631     if (!high_bit_depth && avctx->idct_algo == FF_IDCT_XVIDMMX) {
1632         c->idct_put              = ff_idct_xvid_sse2_put;
1633         c->idct_add              = ff_idct_xvid_sse2_add;
1634         c->idct                  = ff_idct_xvid_sse2;
1635         c->idct_permutation_type = FF_SSE2_IDCT_PERM;
1636     }
1637 #endif /* HAVE_SSE2_INLINE */
1638
1639 #if HAVE_SSE2_EXTERNAL
1640     if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1641         // these functions are slower than mmx on AMD, but faster on Intel
1642         if (!high_bit_depth) {
1643             c->put_pixels_tab[0][0]        = ff_put_pixels16_sse2;
1644             c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_sse2;
1645             c->avg_pixels_tab[0][0]        = ff_avg_pixels16_sse2;
1646         }
1647     }
1648
1649     c->scalarproduct_int16          = ff_scalarproduct_int16_sse2;
1650     c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_sse2;
1651     if (mm_flags & AV_CPU_FLAG_ATOM) {
1652         c->vector_clip_int32 = ff_vector_clip_int32_int_sse2;
1653     } else {
1654         c->vector_clip_int32 = ff_vector_clip_int32_sse2;
1655     }
1656     if (avctx->flags & CODEC_FLAG_BITEXACT) {
1657         c->apply_window_int16 = ff_apply_window_int16_sse2;
1658     } else if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {
1659         c->apply_window_int16 = ff_apply_window_int16_round_sse2;
1660     }
1661     c->bswap_buf = ff_bswap32_buf_sse2;
1662 #endif /* HAVE_SSE2_EXTERNAL */
1663 }
1664
1665 static av_cold void dsputil_init_ssse3(DSPContext *c, AVCodecContext *avctx,
1666                                        int mm_flags)
1667 {
1668 #if HAVE_SSSE3_EXTERNAL
1669     c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_ssse3;
1670     if (mm_flags & AV_CPU_FLAG_SSE4) // not really sse4, just slow on Conroe
1671         c->add_hfyu_left_prediction = ff_add_hfyu_left_prediction_sse4;
1672
1673     if (mm_flags & AV_CPU_FLAG_ATOM)
1674         c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
1675     else
1676         c->apply_window_int16 = ff_apply_window_int16_ssse3;
1677     if (!(mm_flags & (AV_CPU_FLAG_SSE42|AV_CPU_FLAG_3DNOW))) // cachesplit
1678         c->scalarproduct_and_madd_int16 = ff_scalarproduct_and_madd_int16_ssse3;
1679     c->bswap_buf = ff_bswap32_buf_ssse3;
1680 #endif /* HAVE_SSSE3_EXTERNAL */
1681 }
1682
1683 static av_cold void dsputil_init_sse4(DSPContext *c, AVCodecContext *avctx,
1684                                       int mm_flags)
1685 {
1686 #if HAVE_SSE4_EXTERNAL
1687     c->vector_clip_int32 = ff_vector_clip_int32_sse4;
1688 #endif /* HAVE_SSE4_EXTERNAL */
1689 }
1690
1691 av_cold void ff_dsputil_init_mmx(DSPContext *c, AVCodecContext *avctx)
1692 {
1693     int mm_flags = av_get_cpu_flags();
1694
1695 #if HAVE_7REGS && HAVE_INLINE_ASM
1696     if (mm_flags & AV_CPU_FLAG_CMOV)
1697         c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
1698 #endif
1699
1700     if (mm_flags & AV_CPU_FLAG_MMX)
1701         dsputil_init_mmx(c, avctx, mm_flags);
1702
1703     if (mm_flags & AV_CPU_FLAG_MMXEXT)
1704         dsputil_init_mmxext(c, avctx, mm_flags);
1705
1706     if (mm_flags & AV_CPU_FLAG_3DNOW)
1707         dsputil_init_3dnow(c, avctx, mm_flags);
1708
1709     if (mm_flags & AV_CPU_FLAG_SSE)
1710         dsputil_init_sse(c, avctx, mm_flags);
1711
1712     if (mm_flags & AV_CPU_FLAG_SSE2)
1713         dsputil_init_sse2(c, avctx, mm_flags);
1714
1715     if (mm_flags & AV_CPU_FLAG_SSSE3)
1716         dsputil_init_ssse3(c, avctx, mm_flags);
1717
1718     if (mm_flags & AV_CPU_FLAG_SSE4)
1719         dsputil_init_sse4(c, avctx, mm_flags);
1720
1721     if (CONFIG_ENCODERS)
1722         ff_dsputilenc_init_mmx(c, avctx);
1723 }